REVIEW OF MACHINE LEARNING ALGORITHMS IN DETECTING CREDIT CARD FRAUD: TECHNIQUES AND TRENDS

Abstract

Rising digital transactions have heightened credit card fraud, requiring sharper detection methods. Recent works show that performers of 'attention' LSTM networks augmented with reduction techniques like UMAP and PCA, SMOTE data synthesis, and other classifiers arch on the 99% detection accuracy threshold. LSTM models demonstrate predominance over deep structures of opposing nature like autoencoders and CNNs with up to 99.2 precision and 96.3 AUC. Multi-class classifiers resulting from combinations of SMOTE-ENN, SHAP feature engineering, Genetic Algorithms, Random Forests, and ANNs display greater accuracy and precision. Streamlined Automated ML clients, such as Just Add Data, ease endeavors in pipelined set optimization and model refinement. Novel methodologies in the field comprise HGNNs where Attention-temporal decay models and Weighted Attention Ensembles are overlaid. Moreover, more novel works in the field, such as privacy-preserving federated learning enhances detection of credit card fraud on a decentralized, yet secure, collaborative basis.

Keywords (Detecting Credit card Fraud, Long Short-Term Memory (LSTM), Artificial Intelligence, Artificial Neural Learning, Privacy Preserving).